arXiv subjects

Yuxin Cao

Publications and source records attributed to Yuxin Cao.

At least 19 recordsLinked to original sources

arXiv2026

GPU-Resident Speculative Decoding for In-Browser LLM Inference with WebGPU

WebGPU enables private, offline LLM inference inside web pages, but browser dispatch, inter-process communication, synchronous readback and single-token kernels slow decoding and make conventional speculative decoding inefficient. We present a WebGPU runtime that addresses these browser-specific costs. Deterministic small-batch integer kernels accelerate multi-token verification while preserving exact agreement with greedy decoding in the same quantized runtime. A GPU-resident loop performs drafting, verification, acceptance and KV-cache maintenance without a per-token CPU decision. A draft-free context-mixing drafter reuses page content, conversation history and session memory, while a device-aware controller adapts the speculation width and the submission window to measured costs. Across seven models from 0.5B to 8B parameters, 11 tasks and three client GPUs, our runtime decodes 3.5 to 5.4$\times$ faster than WebLLM on a laptop GPU, 4.8 to 6.2$\times$ faster on an integrated GPU and 4.8 to 8.6$\times$ faster on a workstation GPU. The GPU-resident loop and speculation contribute a further 1.8 to 3.2$\times$ over the optimized kernels on the workstation. Outputs match the quantized greedy decoding of the runtime exactly, and the mean GSM8K accuracy matches that of WebLLM. Optional draft heads raise the peak speedup to 10.2$\times$, while session memory provides additional gains without downloading extra weights.

cs.CY↗

arXiv2026

From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents

Coding agents use execution harnesses to constrain the consequences of prompt injection by enforcing permission boundaries. These systems deterministically intercept any out-of-sandbox action and prompt the user for approval. To reduce approval fatigue, users can approve allow-rules proposed by the agent, after which any future action matching the rule is allowed without a prompt. In this work, we show that this mechanism introduces a new risk: untrusted context can influence the permissions presented for approval, with consequences extending beyond the immediate task. We investigate privilege escalation targeting this permission-update process. Specifically, adversarial context planted in a repository can reliably steer an agent performing a routine task into requesting a plausibly justified, yet dangerously broad allow-rule. If the user approves the request, the saved rule can effectively blind the harness, enabling adversaries to execute malicious actions that would otherwise require approval, without triggering human oversight. Across production platforms, including Claude Code and Codex CLI, we demonstrate that agents successfully elicit the target allow-rules in 58% of experimental instances. Once acquired, these rules exhibit a massive semantic blast radius, covering 56.9% of category-specific capability test cases. To further quantify the downstream risk, we show that conditional on rule approval, our exploit achieves verified, silent goal completion in 47 of the 58 eligible instances (81.0%), undermining the safety promise of human-in-the-loop oversight.

cs.AI↗

arXiv2026

VideoSTF: Stress-Testing Output Repetition in Video Large Language Models

Video Large Language Models (VideoLLMs) have achieved strong performance on video understanding tasks, yet existing benchmarks evaluate only what models predict, leaving the stability of how they generate largely unexamined. We surface a previously underexplored generation failure of VideoLLMs, defined as output repetition, in which the decoder collapses into self-reinforcing loops of repeated phrases or sentences, and present VideoSTF, a benchmarking framework for systematically measuring, stress-testing, and exploiting this failure mode. VideoSTF formalizes repetition with three complementary $n$-gram-based metrics, ships a standardized testbed of 10,000 diverse videos, and provides a library of controlled temporal stressors. Across 10 advanced VideoLLMs, VideoSTF reveals four key findings: (i) repetition is pervasive on unperturbed videos and stable across commonly used frame counts, with repetition rates up to 91%; (ii) it spans a severity spectrum from mild redundancy to token-cap loops, and is highly amplified by temporal perturbations; (iii) temporal stressors form a practical black-box attack surface, flipping benign videos into repetitive ones with tens of queries and high attack success rates (up to 98%), and (iv) repetition is not explained by visual redundancy, its amplification tracks local temporal disruption, and only repetition penalties reduce it among common mitigations such as top-$k$ sampling, input filtering, and prompt variation, but increasing the penalty weakens visual grounding. VideoSTF reframes generation stability as a useful and complementary evaluation axis for VideoLLMs and provides the tools to study it. The project page is available at https://videostf.github.io/.

cs.CV↗

arXiv2026

Preserving Provenance in Shared KV Caches for LLM Serving

Production LLM serving stacks combine an inference engine's local prefix cache with a shared KV-cache tier for fleet-wide reuse. The local cache distinguishes requests by adapter, weight configuration and sharing domain, but the shared tier may key entries only by token content and coarse model metadata. This boundary erases provenance and lets identical tokens under incompatible computational or sharing contexts collide. We call this composition gap provenance-blind reuse and present its first systematic study. A source audit of three vLLM connectors confirms the structural omission, while runtime experiments reproduce it across vLLM and two SGLang releases, 12 models from 7 families (0.5 B-32 B), and over 160 configurations. Cross-adapter collisions reduce accuracy from 0.94 to 0.64, incompatible KV representations reduce reasoning accuracy to zero, and salt omission enables 93% prompt identification from timing. We formalize the missing guarantee as the KV provenance contract: for a declared dimension registry, shared keys must be injective over computational and sharing provenance, with identities stable across workers. Any dimension with a stable identity can therefore be added without connector-specific key logic. A canonical descriptor binds per-request and per-worker provenance into lookup and store keys, while a differential checker detects dimensions that change KV state without changing the key. Implemented in vLLM and SGLang 0.5.20 across three cache paths, provenance binding eliminates unsafe reuse while preserving legitimate sharing. Hit-path latency changes remain within 0.34 ms and below run-to-run variation; retention grows with provenance diversity.

cs.DC↗

arXiv2026

Rule-Constrained Assignment for Cue-Ball Identification in Broadcast Snooker

Accurate cue-ball identification is essential for metric analysis of broadcast snooker. Existing systems evaluate each candidate independently against a fixed white prototype and reject candidates above an appearance threshold. Under broadcast conditions, illumination changes can make colored balls appear white, while intrusions from players and equipment can obscure the cue ball or introduce competing candidates. We formulate cue-ball identification as a rule-constrained assignment problem that jointly assigns detected candidates to the bounded snooker inventory: one cue ball, up to 15 reds, and six colors with known spots. The cue ball is selected by the incremental cost of assigning each candidate to the white slot, and the conventional appearance test follows as the one-slot case. On 419 hand-annotated shots, our method improves identity accuracy from 88.1% to 95.5%, and from 80.5% to 95.2% on held-out venues. Within CueLift, our metric state-recovery system, the assignment expands coverage from 36.2% to 55.4% over 6,241 scorable shots. Assigning an estimate to every shot in a separate evaluation on 2,529 shots preserves this advantage.

cs.CV↗

arXiv2026

Runtime Safety Filtering for Two-Terminal Hazards in Robotic Battery Recycling

Runtime safety filters for learned manipulation policies typically define unsafe states as unions of object-wise keep-out regions. This representation can be unnecessarily restrictive for hazards that depend on a joint spatial relation, such as battery recycling, where a conductive payload can short a charged cell only when it approaches both terminals simultaneously. We study runtime filtering for this two-terminal hazard in LIBERO using frozen OpenVLA policies. We factor a runtime filter into three design choices: the predicate structure, its geometric margin, and the fallback action applied when a commanded action is rejected. We compare a conjunctive predicate, a conventional two-site keep-out, and a composite of the two. For each predicate, we vary its margin to obtain a frontier between task success and residual hazard. We then compare four fallback strategies at matched operating points: holding, retreat, sampled search, and a continuous-action barrier projection. Across three workcells, the three predicate families trace nearly identical safety--utility frontiers once each is evaluated over its own margin. In contrast, the fallback strategy has a substantially larger effect: holding reduces task success by up to 0.302 relative to retreat without reducing hazard, while both minimally invasive fallbacks leave substantially more residual hazard. This ordering transfers to a second policy and task suite, while retreat-based filtering remains effective under standing errors in the clearances available to the filter, although correlated error in the estimated payload size is more damaging than larger independent errors in terminal position. These results show that, for proximity-defined manipulation hazards, margin selection and fallback strategy can matter more than predicate structure in determining the safety--utility trade-off of a runtime filter.

cs.RO↗

arXiv2026

Beyond Patch Removal: Persistent Adversarial Effects in Vision-Language-Action Policies

Adversarial patches to Vision-Language-Action (VLA) policies can cause both immediate action corruption and persistent state effects that remain after the patch is removed. Existing evaluations largely focus on continuous attacks and do not separate these two effects. We introduce a state-restoration protocol that removes the patch at matched action-chunk boundaries and measures subsequent recoverability under the same remaining step budget. Clean, random-patch, deviation-matched, and fixed-direction controls distinguish adversarial effects from occlusion, action-error magnitude, and directional persistence. We also evaluate a recovery adapter trained on attack-induced states under controlled intervention latency. On OpenVLA-OFT with EDPA attacks, only 36.2% of LIBERO-Long episodes remain recoverable after five chunks, compared with 89.9% and 87.0% for the deviation-matched and fixed-direction controls. Similar persistent effects are observed on autoregressive OpenVLA. The recovery adapter improves recovery from 7.7% to 47.4% at one-chunk latency, but its benefit decreases substantially with delayed intervention. These results show that adversarial effects can persist after patch removal and that timely intervention is critical for recovery.

cs.CV↗

arXiv2026

COVER: Codec-Robust Video Watermarking with Generative Video Priors

Video watermarking underpins copyright protection and provenance for generated media, yet almost every video is compressed by a codec before it is stored or shared. A codec discards precisely the perceptually redundant components that most watermarks rely on, so the payload is often lost even when the marked video looked flawless beforehand. Existing methods leave this path open, since they treat compression as one entry in a generic list of distortions, while a real codec is not differentiable and cannot enter gradient-based training. We present COVER, the first learned video watermark built around codec compression as its design target, which survives that compression by embedding the payload in the latent space of a frozen generative video autoencoder and recovering it by re-encoding the received video into that same latent space. To make codec robustness trainable, we build a differentiable codec surrogate bank that simulates the dominant degradation modes of practical compression, and we train the embedder and the latent decoder through three shared recovery paths under a fidelity objective that constrains the residual in the pixel and frequency domains. Across four codecs at 12 settings, COVER attains 93.72% average bit accuracy, ranks first on 11 of the 12, improves the strongest prior method by 2.68 points, and lifts the worst operating point from 68.90% to 73.72% while each marked video stays visually close to the source clip that produced it.

cs.CV↗

arXiv2026

Do Input-Level Defenses Transfer to Observation-Level Attacks on VideoLLMs?

Video Large Language Models (VideoLLMs) are increasingly deployed in safety-critical applications such as content moderation and video analytics. To process long videos efficiently, VideoLLMs rely on frame sampling, token compression, and modality fusion, which together form an observation pipeline that reduces the raw video to a compact internal representation. Recent observation-level attacks exploit this pipeline to prevent the model from perceiving harmful content, yet no defense has been explicitly designed for this threat. We introduce DefTEval, a controlled evaluation framework that systematically assesses whether input-level adversarial defenses, which operate on the pixel content of already-sampled frames, can mitigate observation-level attacks. Across five VideoLLMs, eleven representative defenses, and five attack types, we find that input-level defenses offer limited and inconsistent protection, with harmful detection rates frequently near zero. Critically, defenses fail even against attacks that embed harmful signals in every sampled frame, indicating that the bottleneck extends beyond sampling omission to the suppression of signals that do enter the model. Token compression discards localized features, and modality fusion systematically down-weights weakened visual signals. Furthermore, defense effectiveness is dominated by model architecture rather than by the defense method itself, and detection rates vary drastically across content categories, exposing structural weaknesses in temporal reasoning. These findings demonstrate that securing VideoLLMs requires system-level robustness mechanisms spanning sampling-aware coverage guarantees, token-level preservation of safety-relevant features, and modality-balanced fusion.

cs.CV↗

arXiv2026

KeyBound: Keyed and Host-Bound Learned Audio Watermarking for Speech Provenance

Audio watermarking is a proactive route to attributing synthetic speech to its source. Learned audio watermarks are typically judged by payload recovery after a fixed catalog of signal distortions such as noise, compression, filtering, and resampling. That test is necessary but not sufficient for provenance. A mark offered as evidence of origin should not be readable by an unauthorized party, should not be transferable to unrelated audio, and should not vanish when the recording is re-synthesized by a modern generative model. We present KeyBound, a learned audio watermark that restores the two ingredients classical watermarking supplied and learned schemes set aside, a secret key and a host-aware carrier. KeyBound masks the payload with a secret key and embeds the masked bits through a carrier modulated by a frozen spectral representation of the host, so the key governs payload access while the host-conditioned carrier resists direct transplantation. A key-independent presence head lets any party detect a mark, whereas only a key holder reads its attribution, and under the single-sample uniformity assumption a wrong-key decode clears our verification rule with probability at most $2.1\times10^{-3}$. On LibriSpeech against WavMark, AudioSeal, and Timbre, KeyBound holds 1.00 detection accuracy and 0.98 bit accuracy under a spectral denoiser that costs every baseline its detection, decodes at chance without the key, and rejects transplanted carriers. Detection further transfers to held-out DAC and BigVGAN re-synthesis, though exact payload recovery degrades. Speech provenance is thus better posed as a keyed, host-bound attribution problem than as the recovery of a payload under a catalog of signal distortions fixed in advance.

cs.MM↗

arXiv2026

The Temporal Moderation Gap: Text-to-Video Safety Filters Are Blind to Harm in Motion

Text-to-video (T2V) services inherit their safety stack from image generation, pairing a keyword prompt filter with a per-frame checker that blocks a clip whenever one sampled frame looks unsafe. This stack has a blind spot unique to video. We prove that any moderator ignoring frame order accepts a harmful clip whenever it accepts that clip's benign shuffle, so harm carried by the ordering alone escapes. Empirically, the unmodified benchmark prompt already lands a clip in this moderation gap on 32.7% of Sequential-Action targets over four held-out seeds, and paraphrasing, scene splitting, and a feedback-driven prompt search show no significant improvement (paired McNemar $p\ge0.12$), so prompt engineering is not needed to expose the vulnerability. Dense-scoring all 97 rendered frames shows that about a third of the delivered clips merely hide an unsafe frame, while the rest stay harmful as ordered videos even though every frame passes, an order-blind residual the unmodified prompt reaches on a quarter of Sequential-Action targets. We also document a measurement pitfall, since scoring a searched prompt on its own render seed inflates a 7.5% per-generation rate into an apparent 46.7%. A user study confirms that people read these clips as harmful and their shuffles as safe. The fix is to read frame order, and an order-aware detector separates these clips from their own shuffles at AUC 0.74 where per-frame checking sits at chance, which is the signal deployed moderation throws away.

cs.CV↗

arXiv2026

Caved or Convinced: Temporal Sampling Gates Claim Deference in Video Large Language Models

When asked which of two events came first, video large language models can fail in two opposite ways: cave to a false claim, or reject a true one. Prior video sycophancy work measures only the first and mitigates it by teaching the model to trust the user less, a fix known in text and image models to worsen the second. In video, both failures come from two causes the literature treats as one: availability, whether the sparse sampled frames contain the two events, and weighting, whether that evidence is trusted over the user. We separate them with two interventions that keep the claim fixed: a frame-preserving reorder that flips the claim's truth, and a sampling-offset shift that captures or misses both events at a fixed frame budget. When the events are missed, the two twins present identical frames, so each of the nine models we evaluate accepts a true and a false claim at the same rate, making Youden's $J=0$ by construction. Availability is necessary but not sufficient. Five of the nine read the order, yet four of those five still cave to the false claim, so their deference hits a weighting ceiling. Since trust cannot be calibrated over evidence that was never sampled, we propose a reversal test that cancels the model's order prior by scoring the sampled frames forward and reversed, then answers, resamples, or abstains without reading the claim. The test raises the order accuracy to 0.92-1.00 on the models that read the order and abstains rather than guesses on those that cannot.

cs.MM↗

arXiv2026

SpreadMark: Robust Image Watermarking via Spread-Spectrum Embedding

Invisible image watermarks are increasingly used for deepfake detection and provenance tracking, where they must survive not only incidental distortions but also deliberate removal. We revisit spread-spectrum embedding, a classical watermarking principle, inside a modern neural post-hoc watermarking architecture. Our starting point is a measurement: in existing encoder-decoder schemes each message bit occupies only a small fraction of the image, a shared contributing factor to their fragility, since removal then need only disturb the region a bit occupies. SpreadMark instead spreads each bit as a dense pseudo-random codeword over the whole image and recovers it by matched-filtering a learned cover-suppressed chip representation, with a parallel convolutional decoding path and sparsification-aware training. A conditional chip-space analysis shows that, under a codeword-independent perturbation model, dense spreading increases the budget required to disrupt matched-filter recovery. Evaluated on COCO and DIV2K against nine schemes, SpreadMark is the only evaluated method retaining high detection under both the regeneration and the latent-space sparsification settings we test, with competitive JPEG and additive-noise robustness. It keeps the embedded watermark imperceptible, maintaining high perceptual quality on both COCO and DIV2K.

cs.CR↗

arXiv2026

Poisoning Prompt-Guided Sampling in Video Large Language Models

Video Large Language Models (VideoLLMs) are increasingly deployed as automated moderators on user-generated video platforms, where a few unwatched seconds of harmful footage are enough to suppress a safety alert. Because encoding every frame is prohibitive, modern VideoLLMs rely on prompt-guided sampling (PGS), which scores frames against the user prompt and forwards only the top-ranked ones to the visual encoder. Uniform and semantic samplers are known to be defeated by simple frame replacement, whereas PGS, the most prompt-aware family, has escaped scrutiny, and its prompt awareness in fact repairs the omission failures that defeat the other two. We show that this repair is superficial, since PoisonVID, a transfer attack, poisons the sampler's ranking so that harmful clips are never surfaced, without access to target weights, gradients, or sampling internals. It optimizes one video-level perturbation under a relevance-suppression loss defined over a depiction set of paraphrased harmful descriptions written by a shadow VideoLLM and a general-purpose language model, which drives perturbed harmful frames out of the prompt-conditioned subspace that PGS reads. Samplers that never consult that score keep the frames they always kept, which locates the failure at selection rather than at the encoder. Across three PGS methods, six VideoLLMs, and six harmful categories, PoisonVID attains 84% to 97% average attack success over the 18 sampler and model pairs and survives seven defenses. Re-encoding at lower resolution on ingest gives back part of what was evicted and costs the attack 48 points, which bounds the threat without closing it. PGS therefore buys accuracy with a structural safety debt, and sampler design will now have to repay that.

cs.CV↗

arXiv2026

Query-Efficient Video Adversarial Attack with Stylized Logo on Service Computing

In service computing, video classification has become fundamental to many intelligent applications. While Deep Neural Networks (DNNs) have demonstrated excellent performance in recognizing video content, recent studies have shown that DNNs are highly vulnerable to adversarial examples. Thus, understanding adversarial attacks can better respond to emergency situations. In order to improve attack performance, many style-transfer-based attacks and patch-based attacks have been proposed. However, the global perturbation of the former will bring unnatural global colors, while the latter is difficult to achieve success in targeted attacks due to the limited perturbation space. Moreover, compared to a plethora of methods targeting image classifiers, video adversarial attacks remain relatively underexplored. Therefore, to generate adversarial examples with a low budget and to provide them with a higher verisimilitude, we propose a novel black-box video attack framework, called Stylized Logo Attack (SLA). SLA is conducted through three stages. The first stage involves building a style reference set for logos, which can not only make the generated examples more natural, but also carry more target class features in targeted attacks. Then, Reinforcement Learning is employed to determine the style reference and position parameters of the logo within the video, which ensures that the stylized logo is placed in the video with optimal attributes. Finally, perturbations are optimized in a step-by-step manner so as to improve the fooling rate. Experimental results indicate that SLA can achieve better performance than state-of-the-art methods and still maintain good deception effects when facing various defense methods. We believe SLA can raise awareness among the security community about the reliability and security of video classification systems and serve as a memorandum of possible attack methods.

cs.CV↗

arXiv2026

Membership Inference Attacks Against Video Large Language Models

Video large language models (VideoLLMs) are increasingly trained or instruction-tuned on large-scale video--text corpora collected from heterogeneous sources, raising an immediate privacy question: can an external auditor determine whether a particular video was used during training? While membership inference attacks (MIAs) have been studied extensively for classifiers and, more recently, for text and image generation models, the VideoLLM setting remains unexplored. This setting is challenging because black-box auditors observe only generated text, whereas the membership signal is entangled with video-specific factors such as motion complexity and temporal span. In this paper, we present a black-box MIA targeting VideoLLMs that couples temperature-perturbed generation with video-aware difficulty features. Our key intuition is that member samples tend to induce sharper, more brittle generation behavior across decoding temperatures, and that this signal should be interpreted jointly with the intrinsic difficulty of the queried video. Concretely, we query the target model at low and high temperatures, measure the semantic drift between the resulting texts. We evaluate the attack against \texttt{LLaVA-Video-7B-Qwen2-Video-Only} and achieve a member inference AUC of 0.68 and accuracy of 0.63. These results demonstrate that Video-LLMs are vulnerable to black-box membership inference attacks, highlighting an urgent need for the community to systematically evaluate and mitigate privacy risks in VideoLLMs.

cs.CR↗

arXiv2026

DUAP: Dual-task Universal Adversarial Perturbations Against Voice Control Systems

Modern Voice Control Systems (VCS) rely on the collaboration of Automatic Speech Recognition (ASR) and Speaker Recognition (SR) for secure interaction. However, prior adversarial attacks typically target these tasks in isolation, overlooking the coupled decision pipeline in real-world scenarios. Consequently, single-task attacks often fail to pose a practical threat. To fill this gap, we first utilize gradient analysis to reveal that ASR and SR exhibit no inherent conflicts. Building on this, we propose Dual-task Universal Adversarial Perturbation (DUAP). Specifically, DUAP employs a targeted surrogate objective to effectively disrupt ASR transcription and introduces a Dynamic Normalized Ensemble (DNE) strategy to enhance transferability across diverse SR models. Furthermore, we incorporate psychoacoustic masking to ensure perturbation imperceptibility. Extensive evaluations across five ASR and six SR models demonstrate that DUAP achieves high simultaneous attack success rates and superior imperceptibility, significantly outperforming existing single-task baselines.

cs.CR↗

arXiv2026

Towards Stealthy and Effective Backdoor Attacks on Lane Detection: A Naturalistic Data Poisoning Approach

Deep learning-based lane detection (LD) plays a critical role in autonomous driving and advanced driver assistance systems. However, its vulnerability to backdoor attacks presents a significant security concern. Existing backdoor attack methods on LD often exhibit limited practical utility due to the artificial and conspicuous nature of their triggers. To address this limitation and investigate the impact of more ecologically valid backdoor attacks on LD models, we examine the common data poisoning attack and introduce DBALD, a novel diffusion-based data poisoning framework for generating naturalistic backdoor triggers. DBALD comprises two key components: optimal trigger position finding and stealthy trigger generation. Given the insight that attack performance varies depending on the trigger position, we propose a heatmap-based method to identify the optimal trigger location, with gradient analysis to generate attack-specific heatmaps. A region-based editing diffusion process is then applied to synthesize visually plausible triggers within the most susceptible regions identified previously. Furthermore, to ensure scene integrity and stealthy attacks, we introduce two loss strategies: one for preserving lane structure and another for maintaining the consistency of the driving scene. Consequently, compared to existing attack methods, DBALD achieves both a high attack success rate and superior stealthiness. Extensive experiments on 4 mainstream LD models show that DBALD exceeds state-of-the-art methods, with an average success rate improvement of +10.87% and significantly enhanced stealthiness. The experimental results highlight significant practical challenges in ensuring model robustness against real-world backdoor threats in LD.

cs.CR↗